The resource loading... loading...

Ranged Volume DCA Strategy

Author: ChaoZhang, Date: 2023-09-21 10:41:52
Tags:

Overview

This strategy combines the ranged volume indicator and DCA bot strategy, taking positions when ranged volume signals trigger, using DCA bot parameters for pyramiding. It aims to follow trends with low-cost pyramiding.

Strategy Logic

  1. Use ranged volume indicator to identify volume spikes
  2. Go long on spike, pyramid with safety orders when price drops below threshold
  3. Calculate DCA parameters including safety order prices, sizes, max safety orders etc.
  4. Add safety orders when price triggers safety order price levels
  5. Take profit when reaching profit target or max safety orders

Specifically, it combines ranged volume analysis and DCA pyramiding mechanisms. It goes long on volume spike over recent high, and pyramids with safety orders when price drops to each layer. It can track trends but with stop loss limits.

Advantage Analysis

  1. Ranged volume improves entry accuracy
  2. Pyramiding allows low-cost trend following
  3. Flexible configuration of DCA parameters
  4. Take profit and stop loss manage risks

Risk Analysis

  1. Failed volume analysis risks wrong entry
  2. Too many pyramiding increases costs and risks
  3. Requires timely DCA parameter tuning
  4. Improper stop loss placement may expand losses

Risks can be reduced via parameter optimization, adding trend filter etc.

Optimization Directions

  1. Test volume parameter combinations for best setup
  2. Optimize DCA parameters for different products and timeframes
  3. Add trailing stop loss to track price changes
  4. Add re-entry rules to re-enter on trend momentum
  5. Evaluate trend filter to avoid wrong direction entries
  6. Compare stop loss algorithms to find optimal configuration

Summary

This strategy combines ranged volume and DCA mechanisms to enter on volume expansions and pyramid with low cost following trends. Pros are efficient capital use and configurability; Cons are high reliance on parameter optimization. Risks can be reduced through parameter tuning, stop loss optimization while retaining advantages. It allows traders to master using indicators and optimizing trading strategies with bots.


/*backtest
start: 2022-09-20 00:00:00
end: 2023-09-20 00:00:00
period: 1d
basePeriod: 1h
exchanges: [{"eid":"Futures_Binance","currency":"BTC_USDT"}]
args: [["v_input_8",500]]
*/

// This source code is subject to the terms of the Mozilla Public License 2.0 at https://mozilla.org/MPL/2.0/
// © Ranged Volume DCA Strategy - R3c0nTrader ver 2022-04-19
// For backtesting with 3Commas DCA Bot settings
// Thank you "EvoCrypto" for granting me permission to use "Ranged Volume" to create this strategy
// Thank you "junyou0424" for granting me permission to use "DCA Bot with SuperTrend Emulator" which I used for adding bot inputs, calculations, and strategy
//@version=5
strategy("Ranged Volume DCA Strategy - R3c0nTrader", shorttitle="Ranged Vol DCA Strategy", format=format.volume, overlay=true, pyramiding=999, default_qty_type=strategy.cash, initial_capital=50000, commission_value=0.0)

// INPUTS {
// Start and End Dates
i_startTime = input(defval=timestamp('01 Jan 2015 00:00 +0000'), title='Start Time')
i_endTime = input(defval=timestamp('31 Dec 2050 23:59 +0000'), title='End Time')
inDateRange = true

//Ranged Volume Settings
Range_Length    =   input.int(5,        title="Volume Range Length",                       minval=1)

Heikin_Ashi     =   input(true,     title="Heikin Ashi  (Try toggling for different results)")
Display_Bars    =   input(true,     title="Show Bar Colors")
Display_Break   =   input(true,     title="Show Break-Out")
Display_Range   =   input(true,     title="Show Range")

truncate(number, decimals) =>
    factor = math.pow(10, decimals)
    int(number * factor) / factor

// Strategy Inputs
//sourceInput = input.source(close, "Source")
sourceInput = close
price_deviation = input.float(6.0, title='Price deviation to open safety orders (%)', step=0.25, minval=0.0) / 100
take_profit = input.float(22.0, title='Target Take Profit (%)', step=0.5, minval=0.0) / 100
trailing = input.float(0.0, title='Trailing deviation. Default= 0.0 (%)', step=0.5, minval=0.0) / 100
base_order = input(100.0, title='Base order')
safe_order = input(500.0, title='Safety order')
safe_order_volume_scale = input.float(2.0, step=0.5, title='Safety order volume scale')
safe_order_step_scale = input.float(1.4, step=0.1, title='Safety order step scale')
max_safe_order = input(5, title='Max safety orders')

var current_so = 0
var initial_order = 0.0
var previous_high_value = 0.0
var original_ttp_value = 0.0
// Calculate our key levels
take_profit_level = strategy.position_avg_price * (1 + take_profit)

// }


// SETTINGS {
Close = Heikin_Ashi ? request.security(ticker.heikinashi(syminfo.tickerid), timeframe.period, close) : close
//Close = Heikin_Ashi ? request.security(ticker.heikinashi(syminfo.tickerid), timeframe.period, close) : sourceInput
Open = Heikin_Ashi ? request.security(ticker.heikinashi(syminfo.tickerid), timeframe.period, open) : open


Positive        =    volume
Negative        =   -volume

Highest         =   ta.highest(volume, Range_Length)
Lowest          =   ta.lowest(-volume, Range_Length)

Up              =   Highest > Highest[1] and Close > Open
Dn              =   Highest > Highest[1] and Close < Open

Volume_Color = 
 Display_Break and Up ? color.new(#ffeb3b, 20) : 
 Display_Break and Dn ? color.new(#f44336, 20) : 
 Close > Open ? color.new(#00c0ff, 20) : 
 Close < Open ? color.new(#0001f6, 20) : na
// }

//Plot bar color for volume range indicator
barcolor(Volume_Color, title='Ranged Volume Bar Coloring: (You must disable bar coloring in any studies you added or this may not work properly)')
//barcolor(Display_Bars ? Volume_Color : na)

//

// First Position
if strategy.position_size == 0 and sourceInput > 0 and (Up) and inDateRange
    strategy.entry('Long @' + str.tostring(sourceInput)+'💎✋🤚', strategy.long, qty=base_order / sourceInput)
    initial_order := sourceInput
    current_so := 1
    previous_high_value := 0.0
    original_ttp_value := 0
    original_ttp_value

threshold = 0.0

if safe_order_step_scale == 1.0
    threshold := initial_order - initial_order * price_deviation * safe_order_step_scale * current_so
    threshold

else if current_so <= max_safe_order
    threshold := initial_order - initial_order * ((price_deviation * math.pow(safe_order_step_scale, current_so) - price_deviation) / (safe_order_step_scale - 1))
    threshold

else if current_so > max_safe_order
    threshold := initial_order - initial_order * ((price_deviation * math.pow(safe_order_step_scale, max_safe_order) - price_deviation) / (safe_order_step_scale - 1))
    threshold
    

// Average Down
if current_so > 0 and sourceInput <= threshold and current_so <= max_safe_order and previous_high_value == 0.0
    strategy.entry('😨🙏 SO ' + str.tostring(current_so) + '@' + str.tostring(sourceInput), direction=strategy.long, qty=safe_order * math.pow(safe_order_volume_scale, current_so - 1) / sourceInput)
    current_so += 1
    current_so

// Take Profit!
if take_profit_level <= sourceInput and strategy.position_size > 0 or previous_high_value > 0.0
    if trailing > 0.0
        if previous_high_value > 0.0
            if sourceInput >= previous_high_value
                previous_high_value := sourceInput
                previous_high_value
            else
                previous_high_percent = (previous_high_value - original_ttp_value) * 1.0 / original_ttp_value
                current_high_percent = (sourceInput - original_ttp_value) * 1.0 / original_ttp_value
                if previous_high_percent - current_high_percent >= trailing
                    strategy.close_all(comment='Close (trailing) @' + str.tostring(truncate(current_high_percent * 100, 3)) + '%')
                    current_so := 0
                    previous_high_value := 0
                    original_ttp_value := 0
                    original_ttp_value
        else
            previous_high_value := sourceInput
            original_ttp_value := sourceInput
            original_ttp_value
    else
        strategy.close_all(comment='💰 Close @' + str.tostring(sourceInput))
        current_so := 0
        previous_high_value := 0
        original_ttp_value := 0
        original_ttp_value

// Plot TP
plot(strategy.position_size > 0 ? take_profit_level : na, style=plot.style_linebr, color=color.green, linewidth=2, title="Take Profit")

// Plot All Safety Order lines except for last one as bright blue
plot(strategy.position_size > 0 and current_so <= max_safe_order and current_so > 0 ? threshold : na, style=plot.style_linebr, color=color.new(#00ffff,0), linewidth=2, title="Safety Order")

// Plot Last Safety Order Line as Red
plot(strategy.position_size > 0 and current_so > max_safe_order ? threshold : na, style=plot.style_linebr, color=color.red, linewidth=2, title="No Safety Orders Left")

// Plot Average Position Price Line as Orange
plot(strategy.position_size > 0 ? strategy.position_avg_price : na, style=plot.style_linebr, color=color.orange, linewidth=2, title="Avg Position Price")

// Fill TP Area and SO Area
h1 = plot(strategy.position_avg_price, color=color.new(#000000,100), title="Avg Price Plot Area", display=display.none, editable=false)
h2 = plot(take_profit_level, color=color.new(#000000,100), title="Take Profit Plot Area", display=display.none, editable=false)
h3 = plot(threshold, color=color.new(#000000,100), title="SO Plot Area", display=display.none, editable=false)

// TP Area
fill(h1,h2,color=color.new(#38761d,70), title="Take Profit Plot Area")
// Current SO Area
fill(h1,h3,color=color.new(#3d85c6,70), title="SO Plot Area")

More